Illuminating the unseen: Advancing MRI domain generalization through causality.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning excels at accelerated MRI reconstruction but struggles with domain shifts (e.g., varying contrasts, anatomy, acquisition).
• Existing methods lack robustness when applied to unseen MRI datasets, limiting clinical applicability.

Purpose of the Study:

• To develop the first domain generalization framework for accelerated MRI reconstruction.
• To enhance robustness and performance across diverse, unseen MRI domains.

Main Methods:

• Progressive strategies for domain invariance: image-level fidelity consistency and feature alignment.
• Novel mechanism-level invariance enforcement (GenCA-MRI) aligning intrinsic causal relationships.
• Computational strategy to reduce causal alignment complexity for practical use.

Main Results:

• GenCA-MRI demonstrated significant numerical and visual improvements over baseline algorithms.
• Achieved up to 2.15 dB PSNR improvement on fastMRI and 1.24 dB on IXI at 8x acceleration.
• Superior preservation of anatomical details and effective mitigation of domain-shift issues.

Conclusions:

• The proposed domain generalization framework significantly improves accelerated MRI reconstruction robustness.
• GenCA-MRI offers a practical and effective solution for real-world MRI applications facing domain variability.
• This work advances the field by enabling reliable deep learning-based MRI reconstruction across diverse clinical settings.